It's completely on-device [0]. The prediction models have been shrunk [1] so no connections are made (Google has enabled lots of shrunken on-device models with

It's completely on-device [0]. The prediction models have been shrunk [1] so no connections are made (Google has enabled lots of shrunken on-device models with TF Lite [2]) . I'm not sure why this page doesn't call that out considering how privacy-focused the other features are. 0. https://www.androidauthority.com/smart-reply-android-q-983647/ https://www.androidauthority.com/smart-reply-android-q-98364... 1. https://www.tensorflow.org/lite/models/smart_reply/overview https://www.tensorflow.org/lite/models/smart_reply/overview 2. https://arxiv.org/pdf/1708.00630.pdf https://arxiv.org/pdf/1708.00630.pdf, https://www.tensorflow.org/lite/guide https://www.tensorflow.org/lite/guide